Underwater Robotics Training Data

The underwater environment is arguably the most sensor-hostile domain in robotics. GPS does not penetrate water. Radio frequency communication is limited to extremely low frequencies with negligible bandwidth. Visible light attenuates exponentially -- in turbid coastal waters, visibility drops below 1 meter. Pressure increases by 1 atmosphere every 10 meters, reaching over 1,000 atmospheres in the deepest ocean trenches. These fundamental physics constraints make underwater robotics data collection orders of magnitude more difficult and expensive than terrestrial data collection.

Despite these challenges, the underwater robotics market is large and growing. The global subsea robotics market was valued at $4.2 billion in 2023 and is projected to reach $7.3 billion by 2030, driven primarily by offshore energy (oil, gas, and offshore wind), aquaculture, defense, and marine science. The offshore wind industry alone is expected to install over 380 GW of capacity by 2032, and every offshore wind turbine requires regular subsea inspection of its foundations, cables, and scour protection -- tasks increasingly performed by autonomous underwater vehicles (AUVs) and remotely operated vehicles (ROVs).

The leading companies in subsea robotics reflect this market diversity. Oceaneering operates the world's largest ROV fleet with over 250 vehicles serving the offshore energy industry. Saab Seaeye manufactures electric ROVs deployed across energy, defense, and science. Kongsberg Maritime's HUGIN and MUNIN AUVs perform bathymetric survey and pipeline inspection autonomously for days at a time. Blue Robotics has democratized access with affordable ROV platforms for research and inspection. Each deployment requires training data that captures the specific visual and physical conditions of the operational environment.

The data challenge is compounded by the near-total absence of public underwater robotics datasets. While terrestrial robotics benefits from large-scale open datasets (Open X-Embodiment, DROID, nuScenes), no equivalent exists for underwater operations. The underwater domain is data-scarce by nature: every hour of underwater data requires a vessel, an ROV or AUV, trained operators, and favorable sea conditions. This makes every real-world underwater dataset extremely valuable and creates a critical need for purpose-built training data from actual subsea environments.

Oceaneering, the world's largest ROV operator, maintains a fleet of over 250 work-class ROVs deployed globally for offshore oil and gas inspection, maintenance, and construction support. Their Freedom autonomous vehicle performs light intervention tasks without a vessel-deployed tether. Oceaneering's AI ambitions require training data from the enormous diversity of subsea infrastructure they inspect: deepwater wellheads in the Gulf of Mexico, aging North Sea platforms, and new offshore wind foundations in the Baltic and North Seas. The visual appearance of the same structure type varies dramatically by age, marine growth region, and water clarity.

Kongsberg Maritime's HUGIN AUV has completed over 20,000 missions worldwide, performing bathymetric survey, mine countermeasures, and pipeline inspection. HUGIN operates autonomously for 24-72 hours at depths to 6,000 meters, navigating using an integrated INS with GNSS-aided calibration and terrain-relative navigation. The training data for HUGIN's autonomous mission planning comes from thousands of prior surveys that map seabed terrain, current patterns, and acoustic propagation conditions at each operating site.

Saab Seaeye's Sabertooth is a hybrid ROV/AUV that can operate both tethered (for real-time human control) and untethered (for autonomous survey). Equinor uses Sabertooth for resident subsea inspection at the Asgard field in the Norwegian Sea, where the vehicle lives permanently on the seabed in a docking station and performs regular inspections without vessel support. Training data for resident subsea vehicles must capture the gradual changes in a subsea installation over months and years -- progressing marine growth, seasonal visibility variation, and new features like anode wastage.

In the aquaculture sector, companies like BioSort and AKVA Group deploy underwater cameras and ROVs for fish farm monitoring. These systems count fish, estimate biomass, detect sea lice and disease, and inspect net integrity. Norway's aquaculture industry, the world's largest, uses AI-powered fish monitoring across thousands of pens. Training data must include the enormous visual diversity of farmed fish species (Atlantic salmon at various life stages, cod, trout) across seasons, lighting conditions, and water quality.

For defense and port security, companies like ECA Group (now Exail) and Teledyne Marine provide AUVs for mine countermeasures and harbour surveillance. The US Navy's mine countermeasure operations use autonomous sonar-based detection systems that require training data with targets on diverse seabed types: sand, mud, rock, and cluttered environments with natural and man-made debris. The false-alarm rate in mine detection is driven entirely by the quality and diversity of the training data.

Underwater robotics uses a sensor mix dictated by the physics of the ocean. Acoustic sensors dominate at range: multibeam echosounders map seabed bathymetry, side-scan sonar creates acoustic imagery of the seabed and structures, synthetic aperture sonar (SAS) provides high-resolution acoustic images, and acoustic positioning systems (USBL, LBL) provide navigation fixes. Optical sensors are limited to close range: HD cameras (typically 1080p to 4K) with vehicle-mounted LED lighting for visual inspection, structured light systems for 3D measurement of subsea assets, and laser scalers for dimensional reference.

Navigation sensors include inertial navigation systems (INS) with Doppler velocity logs (DVL) for velocity-over-ground measurement, depth sensors (pressure transducers), altimeters for height above seabed, and acoustic positioning receivers. The combination of INS+DVL provides dead-reckoning navigation that drifts at approximately 0.1% of distance traveled, requiring periodic acoustic position fixes.

Emerging modalities include 3D profiling sonar for structure inspection in zero-visibility conditions, laser scanning (LiDAR adapted for underwater use at short range), underwater hyperspectral imaging for biological survey and mineral identification, and chemical sensors (dissolved hydrocarbon, pH, dissolved oxygen) for environmental monitoring. Training data for these emerging sensors is especially scarce, as the technology is still maturing and few operational datasets exist.

Claru collects underwater training data using ROV and diver-operated camera systems across a range of marine environments: offshore platforms, subsea pipelines, port infrastructure, and aquaculture sites. Our collectors include commercial divers and ROV pilots with offshore survival training (BOSIET/HUET certified) who understand the operational context of each inspection task. We capture in conditions ranging from 30-meter tropical visibility to sub-meter turbid North Sea conditions, ensuring training data spans the full operational envelope.

Our annotation pipeline includes subsea-specific protocols: defect classification per API RP 2I and DNV-RP-C203 standards for offshore structural inspection, pipeline condition assessment per DNVGL-RP-F116, marine growth classification by type and severity, cathodic protection assessment from anode imagery, and scour measurement around offshore wind foundations. Our annotators are trained on the visual appearance of underwater corrosion, fatigue cracking, and impact damage, distinguishing real defects from marine growth artifacts that often mimic or obscure damage.

For AUV and autonomous inspection applications, we deliver paired optical-acoustic datasets where camera imagery and sonar data are co-registered to the same subsea features, enabling training of multi-modal fusion models that maintain inspection capability as visibility degrades. All underwater data includes depth, water temperature, visibility estimate, current speed, and positioning metadata. We deliver in industry-standard formats compatible with offshore asset management systems (Veristar, SAP PM, Meridium).